npm / @ai-sdk/google-vertex
@ai-sdk/google-vertex
What changed in @ai-sdk/google-vertex 3 to 5, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.
5 breaking18 new APIs and options16 behavior and defaults18 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli init@ai-sdk/google-vertex 533 changes
Breaking changes 3
Forced tool use is unsupported:
requiredand named tool choices fall back toauto, whilestructuredOutputMode: 'jsonTool'falls back to native structured outputs, each with a warning.Together AI Gemini image inputs are marked unsupported with the current single-image request mapping.
Support for Imagen models and their integration was dropped following the API shutdown.
New APIs and options 12
Adds a provider option and provider metadata to resume responses that used up their maxOutputTokens budget.
Adds gemini-4-argon as a model id.
Exposes the first request's prompt token count in provider metadata.
Adds Claude Haiku 5.5 as a supported model, with a 128k maximum output token limit, adaptive thinking, and five effort levels.
Adds the gemini-nano-banana-2.1 model ID for use with Google models.
Adds experimental streaming lifecycle callbacks for speech generation and transcription telemetry.
Adds the Claude Sonnet 5.5 model ID to @ai-sdk/google-vertex.
Claude Sonnet 5.5 adds the
between_toolsthinking type. Disabled thinking andreasoning: 'none'map to it, budget-based thinking converts to adaptive thinking, andxhighandmaxeffort are lowered tohighwith a warning.Adds Claude Fable 5.1 model IDs to the Vertex provider.
Adds the gemini-3.8-flash model for use with @ai-sdk/google-vertex.
Adds unary transcription via generateContent, with language detection, speaker diarization, word timestamps, and custom vocabulary.
Adds streaming transcription over the Live API WebSocket, with VERBATIM or SMART transcription formatting.
Behavior and default changes 7
Requests that reach the per-request cap (CONTINUATION) are continued automatically, with usage summed across requests.
With the Interactions API, polls background: true model calls until they finish.
Top-level embedding dimensions are mapped to provider settings, allowing providers to receive the requested dimensions.
Thinking can be disabled up to high effort; xhigh and max are lowered to high with a warning when thinking is disabled. Token budgets are converted to adaptive thinking with a warning.
Adds telemetry to speech generation and transcription, with provider usage propagation.
Capability declarations now use confirmed model IDs, leaving unrecognized model names unknown; asynchronous lookups and middleware overrides can also resolve to unknown.
Capability declarations now include single-image editing support for Together AI FLUX.2 Pro and Flex.
Fixes and other changes 11
Fixes surfacing errors that occur during streaming.
Fixes hostname-part validation by using isValidHostnamePart.
Uses a standards-compliant User-Agent header.
Capability declarations now advertise file-input support for QuiverAI Arrow 2 and Arrow 2 Telos.
Supported Vertex tool result URLs are now forwarded as function response file data.
Corrects the advertised supported per-call limit for Gemini images.
Fixes downloading tool result file URLs.
Unclassified empty image results are retried, so these results no longer necessarily end an attempt without retrying.
Google Vertex prompt blocks are marked as terminal.
The default output token setting for Llama 4 was fixed.
Vertex text embedding models now advertise a batch limit of 250.
@ai-sdk/google-vertex 415 changes
Breaking changes 1
Google Vertex content-filtered results are treated as terminal rather than retryable.
New APIs and options 4
Adds the claude-haiku-5-5 model ID to @ai-sdk/google-vertex; the model is recognized with a 128k maximum output token limit and all five effort levels.
Adds the
between_toolsthinking type, the lowest thinking setting onclaude-sonnet-5-5.Adds the gemini-3.8-flash model to the Google provider.
Adds unary transcription via generateContent, with language detection, speaker diarization, word timestamps, custom vocabulary, and VERBATIM or SMART formatting.
Behavior and default changes 7
Automatically continues requests that reach the per-request CONTINUATION cap, sums usage across requests, and exposes the first request's prompt token count.
With the Interactions API, polls background model calls until they finish.
For claude-haiku-5-5, an enabled thinking budget is converted to adaptive thinking with a warning because the model does not support thinking token budgets.
For
claude-sonnet-5-5, budget-based thinking is converted to adaptive thinking.For
claude-sonnet-5-5,structuredOutputMode: 'jsonTool'falls back to native structured outputs with a warning.For
claude-sonnet-5-5, disabled thinking is replaced withbetween_toolsthinking, with a warning.For
claude-sonnet-5-5,requiredand named tool choices fall back toautowith a warning.
Fixes and other changes 3
xhighandmaxeffort are lowered tohighwith a warning because the API rejects them withbetween_tools.Forwards supported GCS tool result URLs as function response file data on Vertex Gemini 3 and later, limited to image, PDF, and text MIME types; HTTP(S) downloads remain supported on all Gemini generations.
Unclassified empty image results are retried, with retry-attempt accounting preserved.
@ai-sdk/google-vertex 39 changes
Breaking changes 1
For claude-sonnet-5-5, disabled thinking is replaced with between_tools thinking, with a warning.
New APIs and options 2
Adds the continuationToken provider option for resuming responses that used up their maxOutputTokens budget.
A provider-independent classification of result retryability is added and preserved through the AI Gateway.
Behavior and default changes 2
For claude-sonnet-5-5, budget-based thinking is converted to adaptive thinking.
Unclassified empty image results are retried, while Google Vertex RAI-filtered results are treated as terminal.
Fixes and other changes 4
Converts enabled thinking with a token budget to adaptive thinking and emits a warning because the model does not support thinking token budgets.
Recognizes the model as having a 128k max output token limit and all five effort levels; when thinking is disabled, xhigh and max effort are lowered to high with a warning.
Diagnostics from completed attempts are preserved.
Google Vertex RAI-filtered results are marked as terminal.